Research On Technology Of Specific Radar Emitter Identification Based On Deep Learning

In modern electronic warfare,electronic counter reconnaissance used to obtain more battlefield intelligence is the first condition to win.It is a powerful method to identify the emitter based on the subtle features of radar signal pulse,and it is also the premise and guarantee to implement accurate jamming in time.The electromagnetic signal environment of modern reconnaissance is complex and changeable.In many cases,the signal-to-noise ratio of intercepted signal is low and the modulation pattern is diverse,but the research of specific radar emitter identification in complex environment is relatively lacking.In order to realize the efficient recognition of radar emitter in complex environment,the feature extraction algorithm of Hilbert pulse envelope using wavelet transform is studied in this thesis,the method of specific radar emitter identification based on recurrent neural network(RNN)which is used to classify the pulse envelope feature waveform and the signal time domain feature waveform respectively is proposed in this thesis,and the experimental results verify the effectiveness of the method.The following is the main work of this thesis: 1.The theoretical basis of deep learning is studied in this thesis,the basic structure of deep learning system is given,the model of convolutional neural network(CNN)and the function of each functional layer are deeply studied.The neural structure of RNN and the basic structure of RNN are analyzed with emphasis on the criteria and methods in the process of neural network training.2.Aiming at the problem that the feature of signal envelope extracted by Hilbert transform is not obvious under low SNR,the wavelet transform is used to reduce the noise of signal envelope,and the simulation experiment is carried out to compare and analyze the two noise reduction methods,sliding window average and wavelet transform.It is found that the signal envelope processed by wavelet transform is more suitable for the original envelope.Considering that the nonlinear Saleh model of power amplifier needs to construct the nonlinear phase change to realize the complexity,the nonlinear Taylor series model of power amplifier is used as the behavior model of power amplifier,and the two models are compared and analyzed through the simulation experiment.It is found that the frequency-domain change of the signal processed by the two models is almost the same,but the Taylor series model of power amplifier is not only simple but also convenient.3.CNN is designed and trained to classify and recognize the pulse envelope characteristic waveform and the signal time-domain characteristic waveform.The simulation results show that the recognition rate of CNN to these two characteristic waveforms is above 98% and above 97% respectively in the environment of low SNR and multi modulation type.In order to further reduce the impact of complex environment on individual recognition,combined with the advantages of RNN in processing time-series data,RNN is designed and trained to classify and recognize pulse envelope characteristic waveform and signal time-domain characteristic waveform.Simulation experiments show that in low SNR and multi modulation environment,the recognition rate of RNN for these two characteristic waveforms is above 99% and above 98% respectively,and the recognition effect is good better than CNN.In the training process of CNN and RNN,the former has fast convergence speed and needs less training set data,but the training time is long,while the latter has slow convergence speed and needs more training data,but the training time is short.